Skip to content

Guides

Bright Data vs Oxylabs for AI answer scraping: engines and cost

Both bill $1.50 per 1,000 AI answers. Oxylabs reaches five engines including Google AI Overviews; Bright Data documents 32 output fields on ChatGPT.

SEO consultant, and more recently GEO, at MentionBox

Bright Data and Oxylabs publish the same entry rate for AI answers, $1.50 per 1,000 results, so the sticker price decides nothing here. The engine list does: Oxylabs can call five of the ten engines we track today and is the only one of the two that returns Google AI Overviews, while Bright Data is down to four after its Perplexity scraper went unavailable on November 10, 2026, and compensates with the wider payload, 32 documented output fields on ChatGPT against Oxylabs' 17 data points, plus 5,000 free records every month. Neither one counts brands, so share of voice, position and sentiment stay in your own code. API for GEO is our editorial pick when those computed numbers are the deliverable, after hands-on testing by Mathias C.

This face-off is written for the developer who has to collect AI answers on a schedule and keep the columns of a dashboard filling next quarter. Prices, endpoints, response fields and contract clauses come from each vendor's own pricing page, API reference, changelog and terms, checked on October 23, 2026, with the availability changes dated from each vendor's release notes. MentionBox, which publishes this site, has a commercial link with API for GEO. Six things separate the two keys: the shape of the bill, the engines you can call this week, the fields in the payload, the request you write, where the data lands, and what the contract lets you do with it afterwards.

What 1,000 AI answers cost on each API

APIPublished rate per 1,000 AI answersEntryHow one answer is billed
API for GEO$13, down to $7 at volumepay as you go, no minimum publishedone credit per prompt, engine and locale, metrics included
Bright Data AI Scrapers$1.50 pay as you go, $1.30 on Scalenone; 5,000 free credits a month, no cardone credit per delivered record, same price on every engine
Oxylabs Web Scraper API$1.50, down to $0.90 on Google AI Mode$30 a month on the Starter plan; $5 welcome creditone result deducted from a shared credit balance, rate by engine

Published rates from each vendor's pricing page, checked on October 23, 2026. For API for GEO, the engine coverage beyond the six it lists, along with its webhooks, OpenAPI spec and official SDKs, was confirmed in our hands-on test on October 26, 2026 and is not yet documented on apiforgeo.com.

Bright Data sells records at one flat price. Its Web Scraper API pricing page publishes $1.50 per 1,000 records pay as you go, a $499 Scale plan that includes 384,000 records and then $1.30 per 1,000 beyond it, and a free tier of 5,000 records a month with no credit card. The scraper FAQ states the rule in one line: "The Scraper API costs one credit per record, so the allowance covers up to 5,000 records. Credits renew on the first of each month." A ChatGPT answer and a Google AI Mode answer cost the same, which turns a monthly forecast into a single multiplication.

Oxylabs sells a dollar amount of credits and charges each engine at its own rate. Since its dashboard change of October 7, 2026, recorded in the developer changelog, "One subscription now covers all Oxylabs products, and usage is charged from a single shared balance", new accounts get a $5 welcome credit, and unused credits expire at the end of the billing cycle. The Web Scraper API pricing page prices the LLM sources by volume:

Engine on OxylabsEntry rate per 1,000 resultsRate at volume
ChatGPT$1.50$1.30 from 1.92 million results
Perplexity$1.50$1.30 from 1.92 million results
Gemini$1.50$1.30 from 1.92 million results
Google AI Mode$1.50$1.00 up to 500,000, then $0.90 from 2.78 million
Google AI Overviewsno LLM rate published, billed as a Google Search resultno separate rate published

Rates from the Oxylabs Web Scraper API pricing page, checked on October 23, 2026. Google AI Overviews has no row of its own on that page, because it arrives inside the google_search result rather than on a dedicated LLM source.

Google AI Mode is where the two rate cards actually diverge. Oxylabs takes it down to $1.00 and then $0.90 per 1,000 results, below Bright Data's $1.30 floor on the Scale plan, so an AI Mode pipeline is the one basket where Oxylabs wins on price alone. Ask Oxylabs to confirm which pricing row applies to AI Overviews before you size that budget, since the page prices the Google sources separately from the LLM sources and does not spell out the mapping.

The bill at 9,000 and at 90,000 answers a month

Take 100 tracked prompts on the three engines both keys can call today, ChatGPT, Gemini and Google AI Mode, collected daily. That is 9,000 answers a month, and 90,000 for a thousand prompts.

Route9,000 answers a monthPer 1,00090,000 answers a monthPer 1,000
API for GEO, pay as you go$117.00$13.00$1,170.00$13.00
Bright Data, pay as you go$6.00, after 5,000 free credits$0.67$127.50$1.42
Oxylabs, credit balance$30.00, the Starter plan floor$3.33$135.00$1.50

Our arithmetic on the rates each vendor published, checked on October 23, 2026, on one stated basket: 100 or 1,000 prompts, three engines, one run a day, no add-ons, the free allowance applied once per month. Change the engine mix and every figure moves.

Four details inside that table decide more than the headline $1.50.

  • The free allowance is worth $7.50 a month, every month. Bright Data's 5,000 credits renew on the first, so a small pilot of 150 prompts a day costs nothing at all, and the first 5,000 records of a bigger month are still free. Oxylabs' $5 welcome credit fires once, and the changelog states that unused free credits "carry over to the first paid billing cycle only".
  • Oxylabs has a floor, Bright Data does not. The cheapest paid Oxylabs plan was $30 a month on the Starter plan when we checked on October 23, 2026, and a month of 9,000 answers consumes $13.50 of that balance. You pay $30 either way, so the effective rate is $3.33 per 1,000 rather than $1.50. Filling the plan is what brings Oxylabs back to its published rate.
  • Unused credits expire. The Oxylabs pricing page states that purchases made through self-service are non-refundable and that unused credits expire at the end of the billing cycle. A quiet month costs the whole plan fee. Bright Data pay as you go and API for GEO both fall to zero when you stop calling.
  • Where Scale pays off. Bright Data's $499 plan equals pay as you go at 332,667 records a month, which is $499 divided by $1.50 per 1,000. Below that volume the plan costs more than the meter.

Which engines each key reaches today

Engine, of the 10 we trackAPI for GEOBright DataOxylabs
ChatGPTyesyes, dataset gd_m7aof0k82r803d5bjmyes, source chatgpt
Perplexityyeslisted, unavailable since November 10, 2026yes, source perplexity
Geminiyes, from our testyes, European countries not availableyes, source gemini
Google AI Modeyesyes, dataset gd_mcswdt6z2elth3zqr2yes, source google_ai_mode
Google AI Overviewsyes, from our testno, served by the separate SERP APIyes, inside the google_search result
Copilotyesyes, named on the product pageno
Grokyeslisted, unavailable since June 14, 2026no
Claudeyesnono
DeepSeekyes, from our testnono
Meta AIyes, from our testnono
Total you can call today1045

Engine lists and availability from each vendor's documentation and release notes, checked on October 23, 2026, with the outage dates as the release notes record them. For API for GEO, ChatGPT, Perplexity, Google AI Mode, Claude, Copilot and Grok are listed on apiforgeo.com, and Gemini, Google AI Overviews, DeepSeek and Meta AI were confirmed in our hands-on test on October 26, 2026.

Three rows in that grid need reading before you promise a dashboard column.

Perplexity is the fresh hole. Bright Data's release notes state that "The Perplexity Scraper is unavailable as of November 10, 2026, and requests currently return no successful results", with no later entry restoring it. Grok has been in the same state since June 14, 2026, after a short return on June 11, and the Grok tool was dropped from version 3.0.0 of the Bright Data MCP server on October 11, 2026. Oxylabs has carried Perplexity as a documented source since 2025 and added a dedicated gemini source on June 23, 2026, so a Perplexity column survives a switch of key.

Copilot runs the other way. Bright Data's AI scraper product page names Bing Copilot among its five engines and links a Microsoft Copilot Scraper API, while its documentation library publishes no field list and no sample payload for it, unlike ChatGPT, Gemini and Google AI Mode. Oxylabs documents no Copilot source at all. Our review of the official Copilot options against scraping sets out what Microsoft itself exposes.

Europe is the gap neither vendor advertises. Bright Data's scraper library marks its Gemini countries list "European countries not available", and the Oxylabs AI Overviews reference names France among the biggest excluded countries for the overview block. A French brand tracking Gemini and AI Overviews is therefore outside both keys' coverage on those two surfaces, which is the sort of thing that reads as a brand losing visibility rather than as a coverage gap.

Claude, DeepSeek and Meta AI are absent from both. An interface scraper reaches what a logged-out browser reaches, and those three sit behind an account. Our ranking of AI answer scraping APIs scores fourteen of these keys on the same four criteria.

Google AI Overviews, on one key of the two

Oxylabs treats the overview as part of a Google search rather than as an LLM target. Its reference asks for source: google_search or google_ads, render: html and a user_agent_type of desktop or mobile, "To maximize the chance of receiving fully rendered AI Overviews content", and the result carries an ai_overviews array, so several AI blocks can come back in one response. The documented fields are specific enough to build a citation report on: answer_text.fragments with a text and a references array holding a source and a url, a bullet_list with its points and their references, a source_panel whose items each carry url, source, title and pos, an info_list, and a pos_overall for where the block sat on the page. A context parameter, expand_aio, expands the block and turns on JavaScript rendering.

Bright Data covers AI Mode instead, and routes the overview to a different product. Its AI visibility page serves Google AI Overview from the SERP API, with its own rate card and its own integration, and its March 2026 release notes add a collapse_aio parameter to the 100-results SERP endpoint, which returns the SERP with the AI Overview section collapsed so that "pixel-position measurements" stay comparable between runs. If Google is the surface you care about most, our face-off on SerpApi and DataForSEO for Google AI Overviews prices the specialists on the same basis, and our ranking of APIs for Google AI Overviews scores the whole field.

What comes back for one ChatGPT prompt

FieldAPI for GEOBright DataOxylabs
Answer textyesanswer_text, answer_html, answer_text_markdown, answer_text_raw, response_rawresponse_text, markdown_text, markdown_json, raw_response
Cited sourcesranked and sorted by categorycitations: url, title, description, iconcitations: title, url, text, description
Search results behind the answernot publishedsearch_sources: url, title, snippet, rank, plus search_sources_morenot published
The search the engine ranown endpointsweb_search_query, web_search_triggerednot published
Links inside the answernot publishedlinks_attached: url, text, positionlinks: url, text
Shoppingyesshopping, shopping_visibleshopping_products with offers and per-product citations
Ads in the answernot publishedadsads with advertiser_info
Maps and local resultsnot publishedis_map, map, business_locationsnot published
Follow-up turn in the same requestnot publishedadditional_prompt, additional_answer_textnot published
When the prompt was sentnot publishedprompt_sent_atnot published
Parser healthnot publishednot publishedparse_status_code, 12000 for success
Brand position, share of voice, sentimentyesnono
Themes and per-brand perceptionyesnono

Field names published in each vendor's API reference, checked on October 23, 2026. Bright Data documents 32 output fields for the ChatGPT Search dataset; the Oxylabs target library counts 17 data points on its ChatGPT card, 24 on Perplexity, 12 on Google AI Mode and 44 on Google AI Overviews.

The shared floor is the answer text and the cited URLs, which is what keeps both prices at $1.50 per 1,000. Above it, Bright Data's ChatGPT scraper documentation lists the three fields we would pick an API for in an audit: search_sources, the results the engine saw with a rank on each, web_search_query, the search it actually ran, and prompt_sent_at, which its March 2026 release notes describe as "the exact time the scraper clicks 'Send' to submit the prompt" rather than the time the API received your request. A log that records the second rather than the first cannot tell a layout change apart from a queue delay.

Oxylabs answers with product detail and with an honest parser signal. Its changelog of October 26, 2026 added the full ChatGPT shopping panel, description, badge, url, merchant, offers[], citations[], image_page_title and image_page_url, with "No add-on or extra parameter is needed", and its entry of July 14, 2026 states that requests with search: true now retry citation extraction automatically, "roughly doubling citation coverage in the parsed citations field". The ChatGPT target reference also carries the warning to read before you compare two runs: the composition of the response varies between a desktop and a mobile query.

Store the citation arrays from the first run. In an analysis we ran in October 2026, the answers that cited anything carried a median of eight cited URLs each, so the per-answer payload is the report, not an accessory to it. Neither API documents a fan-out field, and the closest thing on either key is Bright Data's web_search_query, whose subfields are not published; our explainer on query fan-out in AI search covers how those sub-searches decide which pages get cited, and tracking a brand in ChatGPT with an API sets out the counting code both vendors leave to you, formula by formula.

How to send one prompt on each API

Bright Data is a dataset API. One POST triggers a collection against a dataset ID, and the record comes back from a snapshot or a webhook:

curl -X POST "https://api.brightdata.com/datasets/v3/trigger\
?dataset_id=gd_m7aof0k82r803d5bjm&format=json&include_errors=true" \
  -H "Authorization: Bearer $BRIGHTDATA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"input":[
    {"url":"https://chatgpt.com/","prompt":"best crm for small business",
     "country":"us","web_search":true,"require_sources":true}
  ]}'
{
  "url": "https://chatgpt.com/?q=best%20crm%20for%20small%20business",
  "prompt": "best crm for small business",
  "answer_text": "...",
  "model": null,
  "web_search_triggered": true,
  "citations": [
    { "url": "https://example.com/crm-guide", "title": "...", "description": "...", "icon": "..." }
  ],
  "search_sources": [
    { "url": "https://example.com/crm-guide", "title": "...", "snippet": "...", "rank": 1 }
  ],
  "prompt_sent_at": "2026-09-06T10:22:47.311Z",
  "country": "us"
}

Four inputs are worth setting on every call. url has to be https://chatgpt.com/, prompt tops out at 4,096 characters, web_search defaults to true and is described in the documentation as a permission rather than a guarantee, which is why you read web_search_triggered back, and require_sources returns an error instead of a record when no sources are found. That last one is a billing lever as much as a data one, since Bright Data bills delivered records. The synchronous twin, POST /datasets/v3/scrape, returns the data inline when collection finishes inside a minute and otherwise answers HTTP 202 with a snapshot_id, so one client has to handle both outcomes.

Oxylabs has two request shapes, and they are not interchangeable. The Web Scraper API takes a job on data.oxylabs.io:

curl 'https://data.oxylabs.io/v1/queries' \
  --user "$OXY_USER:$OXY_PASS" \
  -H 'Content-Type: application/json' \
  -d '{
    "source": "chatgpt",
    "prompt": "best crm for small business",
    "search": true,
    "parse": true,
    "geo_location": "United States",
    "callback_url": "https://your-app.example/oxylabs-hook"
  }'

The newer Web API, launched on October 7, 2026, takes the same question under different names:

curl -X POST "https://webapi.oxylabs.io/v1/async/scrape/chatgpt" \
  -H "Authorization: Bearer $OXY_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "best crm for small business",
    "search": true,
    "output": ["json"],
    "location": "US",
    "device": "desktop"
  }'

The field is prompt on one and query on the other, the country is a name on one ("United States") and a two-letter code on the other ("US"), and only the Web API accepts device and a storage block for delivery to S3, Google Cloud Storage, Alibaba Cloud OSS or Tencent TOS. Both paths are asynchronous: the LLM and AI scrapers reference states plainly that for ChatGPT, Gemini and Perplexity "There is no realtime path", so a callback_url or a polling loop is mandatory rather than optional. search exists on ChatGPT only, where false answers from training data with no citations, and Gemini decides for itself whether to browse.

Throughput, batches and where the data lands

  • Published ceilings. Oxylabs prints a figure on its ChatGPT product page, "rate limits up to 100 requests/second", and has accepted up to 5,000 prompts in a single batch request for chatgpt, gemini and perplexity since July 13, 2026, where "Every prompt runs as its own job, so you get one job ID back per prompt". Bright Data publishes no requests-per-second number and advertises unlimited concurrency instead, sizing the job rather than the rate: unlimited inputs on the async trigger, up to 1 GB of input per job, and up to 5,000 target URLs in one call on its product page.
  • Prompt length. 4,096 characters on Bright Data's ChatGPT scraper, 4,000 on the Oxylabs chatgpt source, and 8,000 on google_ai_mode since July 27, 2026. A long prompt set built for one key can need rewriting for the other.
  • File size, which bites at volume. Bright Data caps a delivered file at 5 GB and asks you to set batch_size to the number of records per file; its FAQ warns that a 400 from the delivery endpoint "most often means the batch_size produced a file over 5 GB", and suggests starting around 500,000 records for a payload of roughly 5 KB each. Setting compress: true returns gzip.
  • Retention is the asymmetry. Bright Data keeps a snapshot downloadable for 30 days after the job completes, and splits its webhook in two URLs, notify for the status and endpoint for the records. The Oxylabs LLM documentation publishes no retention window, so your own database has to be the store of record from day one. Both deliver to cloud storage, and Bright Data adds Snowflake, Azure and Pub/Sub alongside S3 and GCS.

Reliability is where neither vendor gives you much to work with. Bright Data's product page advertises a "99.2% Avg. success rate", Oxylabs advertises industry-leading success rates, and both publish comparison pages that measure the other, which is a vendor grading a rival. In our view the only number worth budgeting against is your own: run the same 200 prompts through both keys for a week and count, per engine, the share of requests that came back with a non-empty answer and at least one source.

What you pay for when nothing comes back

At 9,000 answers a month, a 5% empty rate is 450 requests, and the two rate cards treat them almost but not quite the same.

Bright Data charges on delivery. The pricing page promises "Only pay for what's successfully delivered. No hidden fees, no charges for failed deliveries", and the scraper FAQ narrows it: records that fail because the input was incorrect are still billed, "because the request consumed resources". Validate your prompt list and your country codes before a 5,000-input job rather than after. The 5,000 free credits renew on the first and do not accumulate.

Oxylabs also charges on success, stating on its ChatGPT page that "You're only charged for successful results. Failed requests don't consume API credits." Two things still cost money. A parsed response carries a parse_status_code, where 12000 means the parser succeeded and anything else means it failed on some or all fields, so a billed result can arrive with an empty citations array after a ChatGPT layout change. And the money itself expires: unused credits lapse at the end of the billing cycle, and self-serve purchases are non-refundable.

The two changelogs show the same pressure in two forms. Bright Data records outages: Grok in May 2026, back on June 11, gone again on June 14, Perplexity since November 10, 2026. Oxylabs records parser work instead: a consistent request structure across LLM sources on June 23, 2026, citation retries on July 14, custom browser instructions on July 24, granular shopping fields on October 26, and a switch of the ChatGPT ad image field to base64 bytes on October 30. An engine tightening its defences reaches you either as an outage or as a schema change.

What the two contracts allow

Both vendors ground the practice in the same claim, that they read what an ordinary visitor reads. Bright Data states that "all data is collected exclusively from publicly accessible, unauthenticated web interfaces" and argues that "vendor APIs return sanitized 'API-mode' responses that differ from what real users experience". Oxylabs makes the same argument in its own words, that its Web API "simulates real user interactions with ChatGPT UI" and that search: true triggers the engine's own search feature in that interface. Our comparison of the OpenAI web search tool and ChatGPT answers tests that claim against the official route, field by field.

Resale is where both say no.

  • Oxylabs, at clause 4.2.10 of its terms of service, forbids you to "sell, resell (unless explicitly agreed otherwise in the SoW), license, sublicense, distribute, make available, rent or lease the Services", and clause 4.2.9 adds that the services "shall be used only by and for the benefit of the Customer". Two more clauses belong in a procurement review: 4.2.5 bars use for competitive purposes, and 4.3.9 reserves Oxylabs the right to retain data gathered through your use of the Web Scraper API and to use it at its sole discretion. The contracting entity is oxylabs, UAB, of Vilnius, and clause 12.1 puts the agreement under the laws of the Republic of Lithuania.
  • Bright Data, in its service agreement last updated October 12, 2026, requires "Bright Data's prior written authorization" to resell the Service, and forbids you to distribute or sell any Data "in order to offer a similar or competitive product". Resale runs through a partner agreement. Against that, it publishes the compliance file a procurement team asks for: ISO 27001, SOC 2 controls, and GDPR and CCPA statements. The contracting entity is Bright Data Ltd. in Israel, or Bright Data, Inc. for clients in the United States.
  • API for GEO reads the other way, naming resale and white label as an intended use, which is why it tops our ranking of AI visibility APIs for agencies and why our guide on white label GEO data starts there.

A third document belongs in that stack, and it is neither vendor's: each engine publishes its own terms covering automated collection from its interface, and they change. Our reading of what the terms say about scraping ChatGPT goes through them, and your own counsel should see them before a client contract depends on the feed.

Six mistakes to avoid

  1. Reading $1.50 against $1.50 and stopping there. The free allowance, the plan floor and the Google AI Mode tiers move the effective rate by a factor of five at small volume. Write down prompts, engines, countries and runs per day first.
  2. Mixing the two Oxylabs request shapes. prompt and geo_location with a country name belong to data.oxylabs.io, query and location with a two-letter code to webapi.oxylabs.io. Pick one surface per pipeline.
  3. Budgeting Google AI Overviews at the LLM rate on Oxylabs. The overview arrives inside a google_search result with render: html, and the pricing page publishes no LLM row for it. Confirm the rate before you forecast a month.
  4. Keeping a Perplexity or Grok column on Bright Data. Both scrapers are marked unavailable in its release notes, since November 10, 2026 and June 14, 2026. A column that silently stops filling reads as a brand losing visibility.
  5. Promising Gemini in Europe, or AI Overviews in France. Bright Data's Gemini country list excludes European countries, and the Oxylabs AI Overviews reference names France among the biggest excluded countries. Check the country list per engine before you scope a European report.
  6. Letting an Oxylabs result live in the vendor's store. No retention window is published for its LLM sources. Write the payload to your own storage inside the callback handler or the polling loop.

Our view

Three answers, depending on what you are building. Prices checked on October 23, 2026.

  1. Lowest cost per raw answer, and the richest ChatGPT payload. Bright Data AI Scrapers, at $1.50 per 1,000 records pay as you go with no plan fee, $1.30 on the $499 Scale plan past about 332,700 records a month, 5,000 free records renewed monthly, 32 documented output fields including search_sources, web_search_query and prompt_sent_at, 30-day snapshots, delivery into S3 or Snowflake, and the compliance file procurement wants. You pay for it with two engines currently down and no Google AI Overviews. Our score: 6.9 out of 10.
  2. Google AI Overviews from the same key, and the cheaper AI Mode. Oxylabs Web Scraper API, at $1.50 per 1,000 results falling to $1.00 then $0.90 on Google AI Mode, with the overview block fully parsed inside the google_search result, five engines callable today, 100 requests per second published, batches of 5,000 prompts, a parse_status_code that tells you when the parser broke, and cloud delivery to four storage providers. You pay for it with a monthly plan floor, credits that expire, no Copilot and no retention window. Our score: 6.2 out of 10.
  3. Brand figures without writing the analysis. API for GEO, our editorial pick, at $13 per 1,000 answers down to $7 at its largest tier, pay as you go with no published minimum, free reads over answers already collected, resale allowed, and 9.0 out of 10, the best score of the 55 APIs this site rates. In plain terms that is roughly nine times either scraper's rate per answer, and four of its ten engines, Gemini, Google AI Overviews, DeepSeek and Meta AI, plus its webhooks, OpenAPI spec and official SDKs, were confirmed in our test of October 26, 2026 rather than documented on apiforgeo.com.

The reviews go further on each one: Bright Data AI Scrapers, Oxylabs Web Scraper API and API for GEO, with our ranking of self-serve APIs for developers, which weights price and integration at 35% each, and the face-off on cloro and Bright Data if a wider payload per answer is worth more to you than a flat rate.

Frequently asked questions

Is Bright Data or Oxylabs cheaper per 1,000 AI answers?

They publish the same entry rate, $1.50 per 1,000, so volume and the free allowance decide. Bright Data holds $1.50 pay as you go with no plan fee and $1.30 per 1,000 beyond the 384,000 records included in its $499 Scale plan, plus 5,000 free credits renewed monthly. Oxylabs charges $1.50 per 1,000 results on ChatGPT, Perplexity and Gemini, falls to $1.30 at 1.92 million results, and goes down to $1.00 then $0.90 per 1,000 on Google AI Mode. Prices checked on October 23, 2026.

Which AI engines can Bright Data and Oxylabs reach today?

Oxylabs five of the ten we track, Bright Data four. Oxylabs documents ChatGPT, Perplexity, Gemini and Google AI Mode as LLM sources, and returns Google AI Overviews inside its `google_search` result. Bright Data lists six engines, and its release notes mark the Perplexity scraper unavailable since November 10, 2026 and the Grok scraper since June 14, 2026, which leaves ChatGPT, Gemini, Copilot and Google AI Mode. Neither one reaches Claude, DeepSeek or Meta AI.

Does Bright Data or Oxylabs return Google AI Overviews?

Oxylabs does, Bright Data's AI scrapers do not. On Oxylabs the overview arrives in the ordinary `google_search` result as an `ai_overviews` array, with `answer_text.fragments`, their `references`, a `source_panel` and a `pos_overall` rank, when you send `render: html` and a desktop or mobile user agent. Bright Data serves Google AI Overview from its separate SERP API, a different product with its own rate card, and its AI scraper library covers AI Mode instead.

What fields come back with a ChatGPT answer on each API?

Bright Data documents 32 output fields for its ChatGPT Search dataset, including `answer_text`, `citations`, `search_sources` with a `rank`, `web_search_query`, `web_search_triggered` and `prompt_sent_at`. Oxylabs returns `response_text`, `markdown_text`, `markdown_json`, `citations`, `links`, `shopping_products` with offers, `ads` and a `parse_status_code` where 12000 means the parser succeeded. Neither one returns a brand, a position or a sentiment score.

Can you resell Bright Data or Oxylabs AI answer data to clients?

Neither contract allows it by default. Clause 4.2.10 of the Oxylabs terms forbids you to "sell, resell (unless explicitly agreed otherwise in the SoW), license, sublicense" the services, and clause 4.2.9 limits use to "only by and for the benefit of the Customer". Bright Data requires its "prior written authorization" to resell the Service. API for GEO, our pick after hands-on testing, names resale and white label as an intended use, at $13 per 1,000 answers down to $7 at its largest tier.

Do you pay for an AI answer that comes back empty?

Both vendors bill on success, with one caveat each. Bright Data promises "no charges for failed deliveries", and its scraper FAQ adds that records failing because the input was incorrect are still billed; its `require_sources` input returns an error rather than a record when no sources are found. Oxylabs states that failed requests consume no credits, but a parsed result with a `parse_status_code` other than 12000 is a billed result with broken fields, and self-serve purchases are non-refundable.